Updated October 2026
Voice assistants have existed for years.
We can ask them to set timers, play music, check the weather, make calls, or answer simple questions.
But even when the synthetic voice sounds realistic, the interaction often does not feel like a real conversation.
You speak.
The system waits.
It processes the request.
Then it answers.
If you interrupt at the wrong moment, the interaction may become awkward. If the assistant needs to call an external system, you may wait in silence. And if the task requires several steps, the experience quickly stops feeling conversational.
That architecture is beginning to change.
In September 2026, OpenAI released GPT-Live-1 for the API, a full-duplex voice model designed to listen and speak at the same time while delegating deeper reasoning and actions to backend models, tools, or agent frameworks.
Days later, Google made Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking generally available for real-time voice applications. Gemini 3.8 Live is designed for low-latency dialogue, while Extended Thinking adds background reasoning for more difficult, multi-step work.
These releases represent more than improvements in speech synthesis.
They point toward a different interaction model.
Traditional voice AI often resembles:
Listen → Transcribe → Think → Generate → Speak → Wait
The emerging architecture looks more like:
Listen ↔ Speak
while, in parallel:
Understand → Reason → Use Tools → Observe → Adapt
That turns voice AI from a spoken chatbot into something much closer to a live software agent.
The Short Answer
Full-duplex voice AI can continue receiving and interpreting audio while it is producing audio.
In practical terms, the conversation no longer has to behave like a walkie-talkie where one side must finish before the other begins.
Imagine the agent says:
“I found three flights leaving tomorrow—”
You interrupt:
“Only show nonstop flights.”
A capable full-duplex system can recognize that interruption, stop or adjust its response, and continue:
“Sure. There are two nonstop options.”
The system does not have to finish its original speech before responding to the change.
The more important capability appears when voice is combined with tools.
Suppose you say:
“Check my order and move delivery to Friday.”
While the backend retrieves the order, you add:
“Afternoon delivery if possible.”
The system can incorporate the new constraint while the workflow is already running.
That is fundamentally different from the traditional:
speech → text → answer → speech
pipeline.
Table of Contents
Why Traditional Voice AI Often Feels Unnatural
Human conversation is messy.
People pause halfway through sentences. They interrupt each other. They change direction. They hesitate. They say “right,” “okay,” or “mm-hmm” while another person is speaking. Sometimes silence means the speaker is finished; sometimes it simply means they are thinking.
Traditional computer systems prefer something much cleaner:
Input complete
↓
Process
↓
Output complete
↓
Next input
That mismatch is one reason voice assistants can feel robotic even when the generated voice itself sounds excellent.
The problem is not only audio quality.
It is the architecture of the conversation.
Traditional Voice Architecture
A conventional voice assistant often relies on several separate stages.
The user speaks.
A speech-to-text system converts audio into text.
A turn detector decides whether the user has finished speaking.
A language model processes the transcript.
The model generates a text response.
A text-to-speech model converts that response into audio.
The application plays the audio back to the user.
Conceptually:
Speech
↓
Speech-to-Text
↓
Turn Detection
↓
Language Model
↓
Text Response
↓
Text-to-Speech
↓
Playback
This architecture can work extremely well.
But every handoff creates additional engineering complexity.
Latency can accumulate. State needs to remain synchronized. Transcription errors may propagate. Interruption handling becomes harder. Tool calls can freeze the conversation.
OpenAI’s GPT-Live-1 takes a more unified approach to the conversational layer, allowing incoming and outgoing audio to overlap and enabling work to be delegated while the conversation continues.
What Full Duplex Actually Means
The easiest analogy is a phone call.
Two people can technically speak at the same time.
While you are talking, you can still hear the other person.
If they suddenly say:
“Wait—stop.”
you can react immediately.
That is full-duplex communication.
Many older voice systems behave more like walkie-talkies.
One side speaks.
Then the other side speaks.
Full-duplex voice AI allows incoming and outgoing audio streams to coexist.
That supports more natural interruption handling, adaptive pacing, better pause behavior, backchannels, immediate changes of direction, and continued conversation while external work is happening.
Individually, these may look like minor UX improvements.
Together, they can fundamentally change how intelligent the system feels.
Digital Stackroom Framework: Hear → Understand → Reason → Act → Respond → Adapt
A useful way to think about modern voice agents is as a continuous six-stage loop.
1. Hear
The system receives a live audio stream.
It has to deal with speech, silence, pauses, background noise, partial sentences, and interruptions.
2. Understand
The model needs to determine what the user said, what they intend, which context matters, whether they have finished speaking, and whether clarification is required.
3. Reason
The system determines what needs to happen next.
A simple question may require almost no additional reasoning.
A complicated request may need a deeper reasoning model, several tools, or a multi-step workflow.
4. Act
The agent may need to interact with business systems such as calendars, CRMs, databases, booking systems, search tools, payment platforms, internal APIs, or other agents.
5. Respond
The conversational layer keeps the user informed.
That response may be an answer, acknowledgement, clarification, progress update, warning, or confirmation.
6. Adapt
The user may interrupt.
A tool may return an unexpected result.
New information may arrive.
The user’s objective may change.
The system needs to change course without restarting the entire conversation.
The loop becomes:
Hear → Understand → Reason → Act → Respond → Adapt ↺
The important word is continuous.
Modern voice agents are increasingly designed around an interaction loop rather than a sequence of isolated requests.

GPT-Live-1 Separates Conversation From Deeper Work
One of the most interesting architectural ideas behind GPT-Live-1 is delegation.
The voice model is responsible for maintaining the live interaction.
It can listen, speak, handle interruptions, control conversational pacing, and determine when deeper help is required.
More complex work can then be delegated to another model, tool, or agent framework.
Conceptually:
User
↕
GPT-Live-1
Conversation
↕
Backend Agent
Reasoning + tools
↓
Enterprise Systems
This is important because the model that is best at natural, low-latency conversation does not necessarily need to be the model that performs the deepest analysis.
Different parts of the system can specialize.
The Voice Model Does Not Need to Do Everything
Consider this request:
“Compare my current mortgage with three refinancing options and calculate whether switching makes financial sense over five years.”
The conversational model needs to hear the request accurately, maintain context, respond naturally, recognize corrections, and keep the interaction moving.
But the financial calculations and comparison may be better handled by a more capable reasoning model or specialized tool.
A modern system can therefore separate three responsibilities.
Conversation layer: fast interaction, speech, pacing, interruption management.
Reasoning layer: analysis, planning, complex comparison.
Tool layer: retrieval, calculations, APIs, transactions, and other real-world operations.
This modular architecture can make the system both faster and easier to govern.
Gemini 3.8 Live Uses a Related Architecture
Google’s Gemini 3.8 Live is designed as the default option for many low-latency real-time voice-agent experiences.
The current model supports text, images, audio, and video as inputs and can produce text and audio outputs.
It also supports function calling and interleaved reasoning.
A particularly important capability is asynchronous function calling.
A traditional blocking workflow might behave like this:
User asks
↓
Model calls tool
↓
Conversation stops
↓
Tool returns
↓
Model resumes
An asynchronous workflow can behave more like:
User asks
↓
Tool starts
↓
Conversation continues
↓
Tool returns
↓
Model incorporates the result
For voice interaction, that difference matters enormously.
A system that becomes completely silent every time it interacts with another service does not feel like a natural conversational partner.
Extended Thinking Adds Background Reasoning
Not every voice request has the same complexity.
Consider:
“What time is my next meeting?”
That should be fast.
Now consider:
“Compare these three supplier proposals, evaluate cost, implementation risk, and contract terms, then tell me what questions I should ask each vendor.”
That may require deeper reasoning.
Gemini 3.8 Live Extended Thinking is designed for cases where more background reasoning is required during a real-time interaction.
The model can continue processing reasoning and asynchronous tool work while audio interaction remains active.
This suggests an important voice-agent design principle:
conversation and reasoning may operate at different speeds.
Voice AI Needs a Fast Path and a Deep Path
Not every conversational action should trigger maximum reasoning.
A greeting does not need the same computational effort as a complicated financial analysis.
A well-designed voice system can therefore route work through two broad paths.
| Path | Best for | Primary objective |
|---|---|---|
| Fast path | Greetings, confirmations, navigation, simple questions, progress updates | Low latency |
| Deep path | Planning, comparison, multi-step investigation, difficult tool workflows | Higher reasoning quality |
This produces an experience that feels immediate without forcing difficult tasks into a shallow model.
The architecture becomes:
Instant conversation + background intelligence
rather than one model trying to optimize simultaneously for every possible requirement.
Why Latency Matters More in Voice
A three-second pause in a text interface is often acceptable.
A three-second pause after every sentence during a phone call feels uncomfortable.
Voice systems are judged against human conversational timing.
The total delay can include audio capture, networking, model processing, tool execution, reasoning, and audio generation.
Even modest delays at each stage accumulate.
That is why building a production voice agent is not simply:
“Add speech-to-text and text-to-speech to the chatbot.”
The entire system has to be designed around real-time behavior.
Interruption Handling May Matter More Than Perfect Speech
Synthetic voices have become increasingly natural.
But realistic speech alone does not create a good conversational experience.
Imagine an assistant with an excellent human-like voice that continues talking while you repeatedly say:
“Stop.”
The voice may sound impressive.
The interaction is still poor.
A less perfect synthetic voice that immediately recognizes:
“Wait, that’s not what I meant.”
may feel substantially more intelligent.
OpenAI has highlighted interruption handling as one of GPT-Live-1’s core capabilities.
This suggests a broader product principle:
interaction quality is not the same as voice quality.
Turn-Taking Is an Intelligence Problem
Consider:
“I was wondering if…”
Then silence.
Is the user finished?
Probably not.
Now consider:
“Stop. That’s enough.”
The system should respond immediately.
Both situations may contain silence, but they have very different meanings.
Turn-taking depends on language, intonation, context, conversational history, timing, and intent.
The problem is therefore not simply:
“How many milliseconds of silence occurred?”
Increasingly, turn management becomes part of the model’s intelligence.
Backchannels Matter Too
Humans frequently provide small signals while listening:
“Right.”
“Got it.”
“Mm-hmm.”
These are backchannels.
They communicate:
“I’m listening; continue.”
A sophisticated voice agent can use similar signals to make longer conversations feel less mechanical.
This may be particularly useful in tutoring, interviews, support conversations, coaching, and other interactions where one person speaks for an extended period.
But there is a balance.
An agent that says “mm-hmm” every few seconds may quickly become irritating.
Naturalness requires knowing when not to speak.
Voice AI Is Becoming an Orchestration Interface
The most important development may not be the voice itself.
Voice can become the human interface to a much larger agent system.
Imagine saying:
“Move my Monday meeting with Acme to later this week, but don’t schedule anything after 4 PM.”
That may require the system to identify the meeting, retrieve participant information, inspect calendars, apply time constraints, find possible alternatives, update the calendar, and confirm the result.
The spoken conversation is only the front door.
Behind it is an orchestration workflow.
That changes the role of voice from:
input modality
to:
agent interface
Separate the Conversation Layer From the Workflow Layer
This distinction becomes especially important in production systems.
The conversation layer should focus on communication.
It manages tone, pacing, turn-taking, clarification, interruptions, acknowledgements, and progress updates.
The workflow layer should focus on business execution.
It handles application state, APIs, authentication, permissions, transactions, retries, and deterministic business rules.
A stronger architecture therefore looks like:
Voice model
↓
Agent orchestrator
↓
Identity and policy layer
↓
Tools and APIs
↓
Business systems
↓
Audit and monitoring
Separating conversational intelligence from business authority makes systems easier to test and govern.
OpenAI GPT-Live-1 vs Gemini 3.8 Live
These systems are evolving quickly, but as of October 2026 the broad comparison looks like this:
| Capability | GPT-Live-1 | Gemini 3.8 Live |
|---|---|---|
| Primary focus | Natural full-duplex voice conversation | Low-latency real-time multimodal interaction |
| Listen and speak concurrently | Yes | Real-time bidirectional interaction |
| Audio input/output | Yes | Yes |
| Text input | Yes | Yes |
| Image/video input | Check current API modality support for your implementation | Yes |
| Tool use | Backend delegation | Function calling |
| Async workflow model | Delegation to backend agent/tools | Non-blocking asynchronous function calling |
| Deeper reasoning | Delegate to backend model | Interleaved reasoning |
| Separate high-reasoning live model | Backend model choice | Gemini 3.8 Live Extended Thinking |
The more important point is not which architecture “wins.”
Both illustrate the same broader shift:
live conversation does not need to stop while software work happens elsewhere.
Voice Agents Can Act While Talking
Consider a hotel booking.
A rigid voice bot may ask one question at a time:
“Which city?”
Then:
“Which date?”
Then:
“How many guests?”
A more capable voice agent can process something like:
“I need a hotel near HITEC City in Hyderabad next Wednesday for two nights. Keep it under ₹8,000 per night, and I need reliable Wi-Fi.”
The system can extract multiple constraints immediately.
While the booking service searches, the agent might ask:
“Would you prefer to be within walking distance of restaurants, or is being closer to the office more important?”
The workflow continues while the conversation refines the goal.
This is closer to how humans naturally collaborate.
Voice Agents Need Stronger Confirmation Rules
Voice makes interaction easy.
That is an advantage.
It can also make consequential actions feel too easy.
Consider:
“Cancel my reservation.”
“Send the payment.”
“Delete that record.”
“Share that document.”
A natural voice interface should not make irreversible actions invisible.
A safer pattern is:
Request → Understand → Show or speak intended action → Confirm → Execute → Audit
For high-risk operations, the system may need stronger authentication or a secondary approval channel.
Conversational convenience should not eliminate deliberate control.
Digital Stackroom Voice-Agent Risk Model
A useful way to determine how much autonomy a voice agent should receive is to combine consequence with reversibility.
| Type of action | Example | Recommended control |
|---|---|---|
| Low impact, reversible | Search information, summarize records | High automation |
| Moderate impact, reversible | Create appointment draft, prepare email | Automate with confirmation |
| Moderate impact, external | Send customer message, change booking | Explicit approval |
| High impact | Payment, account change, sensitive disclosure | Strong authentication + approval |
| High impact, difficult to reverse | Delete data, alter permissions, production action | Strict controls and audit |
The more consequential and irreversible the action, the less the system should rely on conversational intent alone.
Identity Becomes Critical
When a voice assistant merely answers general questions, identity may not matter much.
When the same assistant can access calendars, customer records, payments, internal documents, or enterprise systems, identity becomes central.
The system needs to establish:
Who is speaking?
Who is the agent acting for?
What is that person allowed to do?
Which action is being authorized?
Voice recognition alone should not automatically be treated as strong authorization for sensitive actions.
Voice-agent security needs to be built around identity and policy, not simply natural conversation.
Privacy Is More Complicated With Always-Available Audio
Voice systems process a form of data that is inherently personal.
Audio can contain names, account information, private conversations, background speech, health information, payment information, or other sensitive context.
Organizations deploying voice agents therefore need clear policies around recording, retention, transcription, consent, redaction, access, and deletion.
The most natural agent should not become the least transparent system in the organization.
Users need to understand when audio is being processed and what happens to it.
Voice Agents Need Observability
A text-based agent can already be difficult to debug.
Voice adds another layer.
A successful evaluation needs to look beyond the final spoken response.
The organization may need to inspect:
What did the user say?
What did the model understand?
When did it interrupt?
What work was delegated?
Which tools were called?
What did those tools return?
Which state changed?
What did the agent finally communicate?
OpenAI’s evaluation guidance for GPT-Live emphasizes this end-to-end view because a natural-sounding confirmation does not prove that the correct backend action actually happened.
For production voice agents, “it sounded good” is not a sufficient test.
The Voice-Agent Evaluation Scorecard
A practical evaluation should cover multiple dimensions.
| Dimension | Example question |
|---|---|
| Conversation | Did the system respond naturally? |
| Turn-taking | Did it wait when appropriate and interrupt when necessary? |
| Understanding | Did it interpret the user’s intent correctly? |
| Tool execution | Were the correct tools and arguments used? |
| Task completion | Did the real-world state change correctly? |
| Recovery | Could the system handle corrections and failures? |
| Safety | Were sensitive actions appropriately confirmed? |
| Latency | Did the conversation remain responsive? |
| Transparency | Did the user know what the agent was doing? |
A production voice agent needs to perform well across the entire system, not only on speech quality.
Voice and Screens Will Work Together
Voice is excellent for some tasks.
It is useful while walking, driving, working with your hands, or when rapid conversational interaction matters.
Screens remain better for detailed tables, visual comparison, dashboards, long documents, editing, maps, code, and dense information.
The likely future is therefore not:
voice replacing graphical interfaces.
It is:
voice + screens + agents.
You may begin a task through voice, allow an agent to do background work, then review the result visually.
The modalities complement each other.
Voice Agents + Data Agents
One particularly powerful combination is voice plus analytical agents.
Imagine asking:
“Why did revenue decline yesterday?”
The voice system maintains the conversation.
A data agent investigates.
You interrupt:
“Focus only on APAC.”
The analytical workflow changes direction.
The backend returns evidence.
The voice layer explains the main finding and offers to display the detailed breakdown on screen.
That creates a new analytical interaction model:
Ask → Investigate → Interrupt → Refine → Explain → Visualize
Voice becomes the conversational interface.
The data agent handles investigation.
The screen handles dense evidence.
Voice Agents + Workspace Agents
The same pattern can apply to everyday knowledge work.
Imagine asking while commuting:
“What do I need to know before my first meeting?”
A permitted agent could inspect your calendar, relevant email, recent documents, and project context.
The voice layer then gives you a briefing.
You respond:
“What was the unresolved issue from last week’s discussion?”
The system retrieves more context and continues.
This begins to look less like a voice assistant and more like a persistent work interface.
Voice Agents + Tool Protocols
As agent ecosystems become more standardized, voice can sit above a broad catalog of tools.
Conceptually:
Human voice
↓
Full-duplex voice model
↓
Agent orchestrator
↓
Tool layer
↓
Enterprise APIs
The voice system does not need custom conversational logic for every possible enterprise action.
Instead, it can operate as the human-facing layer of a larger tool ecosystem.
That could make voice far more useful than previous generations of assistants.
Where Voice Agents May Deliver the Most Value
The best use cases share one characteristic:
speaking is materially easier than navigating software.
Examples include customer service, field operations, logistics, travel, accessibility, hands-busy environments, sales assistance, language learning, interview support, and some forms of personal productivity.
Voice is less attractive when the task depends heavily on complex visual comparison or when privacy makes speaking aloud inappropriate.
The product question should therefore not be:
“Can we add voice?”
It should be:
“Which task becomes substantially easier because voice exists?”
A Practical Adoption Roadmap
Organizations should introduce voice-agent autonomy gradually.
Phase 1 — Information
Start with answering questions, searching approved knowledge, summarizing information, and explaining processes.
Phase 2 — Read-Only Personalization
Allow access to permitted calendars, account information, project context, or customer records without modifying anything.
Phase 3 — Reversible Actions
Introduce actions such as preparing drafts, creating tentative appointments, or updating low-risk internal records.
Phase 4 — Controlled Transactions
Allow higher-impact actions only with explicit confirmation, strong permissions, and reliable audit trails.
Phase 5 — Complex Agentic Work
Only after the previous layers are well tested should the system autonomously coordinate multi-step workflows across several tools.
The goal is not maximum autonomy.
It is appropriate autonomy.
What Developers Need to Learn
Voice agents are systems-engineering projects.
Prompting matters, but production systems also require expertise in streaming audio, WebSockets or WebRTC, latency management, turn detection, interruption recovery, session state, function calling, orchestration, authentication, observability, error recovery, and testing.
Google’s Live API, for example, uses persistent real-time connections and streams audio and other realtime inputs throughout the session.
The quality of the final product depends on the entire pipeline.
What Product Teams Need to Learn
Natural speech creates strong expectations.
When an agent sounds fluent and human-like, users may assume it is equally strong at reasoning, memory, emotional understanding, or real-world execution.
That assumption can be dangerous.
Product teams should clearly communicate the system’s authority and limitations.
A voice agent should never use natural speech to hide uncertainty.
If the backend cannot verify an action, it should say so.
Instead of:
“Your account was updated.”
when the system is uncertain, it should say:
“I haven’t been able to confirm the update yet. I’m checking the status again.”
Confidence calibration is part of voice UX.
The New Interface for Agentic Computing
Software interaction has evolved through several dominant models.
Graphical applications required users to navigate screens.
Chat interfaces allowed users to describe what they wanted.
Agents began executing multi-step workflows.
Full-duplex voice adds another layer:
users can describe, interrupt, refine, and redirect a goal while the workflow is running.
That is a meaningful shift.
Speech is one of the most natural interfaces humans have.
Historically, computers were simply not good enough at participating in conversation.
Full-duplex systems are beginning to reduce that gap.
Will Voice Replace Apps?
No.
Voice is poorly suited to many tasks.
Try comparing 40 rows of financial data verbally.
Or editing a detailed presentation without seeing the slides.
Or reviewing code purely through speech.
The more realistic future is multimodal.
You may say:
“Compare our three strongest markets.”
The agent performs the analysis.
The screen displays a chart.
You ask:
“Why did India improve?”
The agent investigates and explains verbally.
You then say:
“Add that breakdown to the report.”
Voice provides intent.
Agents coordinate work.
Screens provide visual precision.
Frequently Asked Questions
What is full-duplex voice AI?
Full-duplex voice AI can receive and interpret incoming audio while also generating outgoing audio, allowing more natural interruptions, overlap, and conversational turn-taking than strict one-speaker-at-a-time systems.
What is GPT-Live-1?
GPT-Live-1 is OpenAI’s full-duplex API voice model designed for natural, expressive real-time conversations with smooth interruption handling. It can delegate deeper reasoning and tool use to backend models or agents.
When was GPT-Live-1 released?
OpenAI released GPT-Live-1 in the API on September 10, 2026.
How much does GPT-Live-1 cost?
OpenAI currently lists voice sessions at $0.05 per minute, billed per second. Backend model and tool usage is billed separately.
What is Gemini 3.8 Live?
Gemini 3.8 Live is Google’s real-time low-latency model for voice-agent and Live API experiences. It supports text, image, audio, and video inputs, text and audio outputs, interleaved reasoning, and function calling.
What is Gemini 3.8 Live Extended Thinking?
Gemini 3.8 Live Extended Thinking is Google’s higher-reasoning Live model for complex, multi-step real-time voice tasks. It can perform background reasoning and asynchronous tool work while maintaining a live interaction.
Can voice agents call external APIs?
Yes. Modern voice-agent architectures can connect to tools, functions, APIs, databases, enterprise systems, and backend agents. The exact capabilities depend on the platform and implementation.
Can users interrupt full-duplex AI?
Yes. Interruption handling is one of the central advantages of full-duplex interaction.
Is voice AI safe for financial or other high-impact transactions?
Voice can be part of a high-impact workflow, but sensitive actions should use strong identity controls, explicit confirmation, policy enforcement, and auditing rather than relying on conversational intent alone.
Will voice replace graphical interfaces?
Unlikely. Voice is better suited to some situations, while screens remain superior for dense visual information and precise editing. The more likely future combines voice, graphical interfaces, and agents.
Final Thoughts
The most important development in full-duplex voice AI is not that synthetic speech sounds more human.
It is that the interaction architecture is becoming more human.
Traditional voice systems behave approximately like:
Listen → Wait → Think → Speak
Modern full-duplex systems increasingly behave like:
Listen ↔ Speak
while simultaneously:
Reason → Use Tools → Observe → Adapt
That changes what voice software can become.
A voice interface no longer needs to be limited to answering simple questions.
It can become the conversational front end to enterprise data, productivity software, analytics, customer-service platforms, booking systems, workflow automation, and AI agents.
OpenAI’s GPT-Live-1 demonstrates an architecture built around fast full-duplex interaction with delegated backend intelligence.
Google’s Gemini 3.8 Live family demonstrates real-time multimodal interaction combined with interleaved or background reasoning and asynchronous tool execution.
Different architectures may win in different scenarios.
But the broader direction is clear.
Voice is moving from:
a way to talk to software
toward:
a way to collaborate with software while it works.
And that may make full-duplex voice one of the most important interfaces for the next generation of agentic computing.
Official Sources & Further Reading
OpenAI — Build More Natural Voice Experiences with GPT-Live-1 in the API
Official September 2026 launch announcement covering full-duplex conversation, interruption handling, delegation, customization, and voice-agent architecture.
OpenAI — GPT-Live-1 Model Documentation
Current API model documentation covering model capabilities, backend delegation, and pricing.
OpenAI — GPT-Live Evaluation Guide
Guidance for evaluating conversation quality, delegated work, tool execution, resulting application state, and the final response communicated to the user.
Google AI for Developers — Gemini 3.8 Live
Current model documentation covering supported modalities, interleaved reasoning, asynchronous function calling, and Live API support.
Google AI for Developers — Gemini 3.8 Live Extended Thinking
Documentation for Google’s higher-reasoning real-time model with background reasoning and asynchronous tool work.
Google AI for Developers — Live API
Current documentation for real-time bidirectional audio, video, text input, streaming audio output, and live tool integrations.

